Download files
Abstract
This thesis develops new automatic methods to improve identification, segmentation and analysis of the pulmonary bronchovascular anatomy (PBVA) on computed tomography (CT) images in order to enable more precise detection and monitoring of diseases that impact on these anatomical structures. This is useful because the PBVA is subject to morphological change in some pathological conditions. The arborising nature of PBVA and its close relationship to adjacent structures, and thin bronchial walls make segmentation and analysis difficult.
To automatically identify the main pulmonary artery (MPA) various machine learned models are applied to pre-processed CT images. SVMs demonstrate the highest accuracy (>94%) of MPA detection when compared to ANN, NBC and kNN models. An algorithm is developed that segments and analyses MPA morphology and demonstrates accurate measurements when compared to those performed manually by a radiologist over 24 test CT data sets. Techniques for peripheral pulmonary vasculature segmentation and analysis are implemented and help differentiate pulmonary arteries from veins and in the detection of peripheral bronchi. Tree data structures of the segmented vasculature are created and cross-sectional images perpendicular to vessel centrelines are constructed, enabling analysis of the properties of vascular branches, including the presence of an accompanying bronchus.
A novel technique for segmenting the proximal bronchial tree using isosurface projection is described. Choosing isosurfaces with isovalues close to air (specific for the central bronchial tree) and mapping to those of slightly higher isovalues (accurate for airway inner walls) enables accurate segmentation of the airway down to the lobar level and into most segmental bronchial branches. By applying a novel ring filter and hierarchical clustering to the cross-sectional images of the vascular branches, small to medium diameter bronchial segments are identified with high precision. Optimal parameters for the ring filter and hierarchical clustering algorithm achieve bronchial detection precisions of over 95% in 21 CT test data sets.
Future work on peripheral bronchial identification will lead to accurate segmentation of the peripheral airway, which can then be utilised to separate the pulmonary arterial and venous systems. Complete segmentation of the PBVA would then allow comprehensive computerised analysis of the relationship between PBVA morphology and pathological processes.